EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Real‐Time Federated Fault Detection for Substations With Incomplete Multimodal Data

Xiangyu Zhao, Shang Cao, Lingyu Liang, Jingxing Yu, Huanming Zhang, Ruiye Zhou, Ge Ma

The Journal of Engineering · 2026-01 · 2026 (1)

DOI: 10.1049/tje2.70222 ↗

Kurz gefasst

Fachartikel in der Zeitschrift The Journal of Engineering, online erschienen am 24.08.2026. Verfasst von Xiangyu Zhao und 6 weiteren Autor:innen. Die Arbeit nennt 33 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: China Southern Power Grid.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Substation fault detection models are increasingly being deployed on edge devices for vibration analysis and image classification. Existing methods require multiple edge nodes to share raw data, leading to privacy concerns. Privacy‐preserving federated architectures typically rely on single‐modal approaches, which are insufficient under extreme conditions such as intense light, smog, or darkness. This results in the underutilisation of the complementary information offered by different modalities. Infrared images can reveal thermal anomalies under extreme conditions, while visible light images can capture surface details that infrared images cannot distinguish, thereby improving the model's robustness. In this paper, we present federated edge diagnosis with multimodal fusion data (FED‐MFD), a distributed real‐time fault diagnosis framework that transforms cross‐modal perceptual cues into privacy‐preserving, edge‐executable detection. FED‐MFD extracts semantic features from infrared and visible‐light streams using a dual‐branch feature extraction network (DBFEN), aligns heterogeneous representations via a bimodal regional attention mechanism (BRAM), and performs inference through a lightweight detection head (LDH) designed for resource‐constrained devices. To address client‐side modality absence and model heterogeneity, the federated layer adaptively freezes parameters from the missing modality, distills bimodal knowledge into unimodal clients, synthesises pseudo‐features using a Modality Hallucination Module, and aggregates the results using a quality scoring mechanism sensitive to modality integrity and detection accuracy. Extensive experiments on a real‐world dataset from five substations demonstrate that FED‐MFD achieves 91.8% mAP@0.5, 63.5 FPS and 32.4 million parameters, outperforming state‐of‐the‐art federated and multimodal baselines.

Förderung

  • China Southern Power Grid

Zitieren

APA

Zhao, X., Cao, S., Liang, L., Yu, J., Zhang, H., Zhou, R., & Ma, G. (2026). Real‐Time Federated Fault Detection for Substations With Incomplete Multimodal Data. The Journal of Engineering, 2026(1). https://doi.org/10.1049/tje2.70222

Vancouver

Zhao X, Cao S, Liang L, Yu J, Zhang H, Zhou R, et al. Real‐Time Federated Fault Detection for Substations With Incomplete Multimodal Data. The Journal of Engineering. 2026;2026(1). doi:10.1049/tje2.70222

BibTeX
@article{Zhao2026,
  author = {Zhao, Xiangyu and Cao, Shang and Liang, Lingyu and Yu, Jingxing and Zhang, Huanming and Zhou, Ruiye and Ma, Ge},
  title = {Real‐Time Federated Fault Detection for Substations With Incomplete Multimodal Data},
  journal = {The Journal of Engineering},
  year = {2026},
  volume = {2026},
  number = {1},
  doi = {10.1049/tje2.70222}
}
Kostenlos starten →